New info-theoretic model unifies navigation in biology and AI

Navigating to find resources is a fundamental challenge for both biological organisms and artificial agents. A new paper from Yale and University of Chicago researchers introduces an information-theoretic framework that quantifies the tight coupling between sensory inputs and behavioral actions using transfer entropy. They decompose navigation strategies into two components: a reactive component measuring information flow from sensory inputs to behavior, and an active component measuring how behavior shapes subsequent sensory experiences. This bidirectional information loop, they argue, is a universal driver of sensory navigation.

Applying this framework to experimentally measured trajectories of E. coli bacteria, C. elegans worms, Drosophila flies, and a reinforcement-trained agent navigating a sensory landscape, the team found that bidirectional information flow consistently predicts navigation efficiency. The decomposition exposed distinct strategies: bacterial chemotaxis relies heavily on reactive control, while fly olfactory navigation shows spatial dependency in the active-reactive balance. The RL agent learned policies that mirrored biological strategies. This work establishes a unifying principle that could guide the design of more efficient autonomous navigation systems.